Two Routes to More Capable LLMs: Post-Training Extraction vs. Retraining the Foundation Model

    This article compares two competing routes for making large language models more capable: pushing an existing base model to its ceiling through post-training, and rebuilding a larger foundation model from scratch. It breaks down both approaches across three dimensions—underlying principles, cost, and capability limits—covering supervised fine-tuning, preference alignment (RLHF/DPO), verifiable-reward reinforcement learning, MoE sparse architectures, and data and training strategy. It also examines how each route affects coding and agentic capabilities, and offers a decision framework for engineering teams choosing between fine-tuning an existing model and retraining a new one.

    Introduction: A debate over “where capability comes from”

    In the current wave of model development, you often see two technical narratives that look contradictory but each hold up on their own. One emphasizes leaving the underlying base untouched and squeezing the model’s existing intelligence to its limit through careful post-training. The other argues for building a new, larger base, pulling ahead from the ground up with a sparser expert structure and far more data. Both target the same goal—“stronger coding and agent capabilities”—yet they represent two very different logics of investment and return.

    Understanding these two routes is not about picking a side. It is about giving engineering teams a clear framework for making real decisions like “keep fine-tuning the current model” versus “train a new model from scratch.” Below, I break down each route across three dimensions: principles, cost, and capability limits.

    1. Route One: Keep the base, squeeze intelligence to its ceiling through post-training

    What “post-training” means

    Training a large model is usually split into two stages. Pre-training learns language and world knowledge from massive unlabeled corpora, and it determines the model’s “knowledge foundation” and its upper bound. Post-training then takes the base produced by pre-training and, using a relatively small but high-quality dataset, aligns the model to human intent, specific task formats, and reasoning patterns.

    The core belief of the post-training route is that a sufficiently strong base often leaves its latent capability untapped. With high-quality supervision and reinforcement learning, you can approach—or even push—the base’s “intelligence ceiling.”

    Key stages of post-training

    A typical, industry-standard post-training pipeline usually looks like the illustrative pseudocode below:

    1. # Post-training pipeline illustration (not a specific implementation)
    2. base_model = load_pretrained_checkpoint()
    3. # 1) Supervised fine-tuning (SFT): establish basic alignment with high-quality instruction-response pairs
    4. sft_model = supervised_finetune(base_model, curated_instruction_dataset)
    5. # 2) Preference alignment: make outputs better match human preferences
    6. # Common approaches include RLHF, DPO, and other preference optimization
    7. # methods that need no explicit reward model
    8. aligned_model = preference_optimization(sft_model, preference_pairs)
    9. # 3) Reasoning enhancement: reinforcement learning on verifiable tasks (e.g., math, code)
    10. reasoning_model = verifiable_reward_rl(aligned_model, reward=executor_or_checker)
    11. return reasoning_model

    Three of these stages are worth expanding on:

    • Supervised fine-tuning (SFT) determines “what format and style the model answers in.” Data quality matters far more than quantity; a small set of carefully labeled data often beats a large set of noisy data.
    • Preference alignment (RLHF / DPO-type methods) optimizes the model over the distribution of outputs that humans prefer. Methods like DPO skip the separate step of training a reward model, making them lighter-weight from an engineering standpoint and well suited to resource-constrained teams.
    • Reasoning enhancement (reinforcement learning with verifiable rewards) is a key technique for unlocking reasoning ability in recent years. For tasks with objective right-or-wrong answers such as code and math, you can use test-case pass rates or answer checkers as reward signals, letting the model explore better solution paths on its own.

    Advantages and the ceiling

    The biggest advantage of the post-training route is controllable cost and fast iteration: with no need to rerun expensive pre-training, a team can produce new versions quickly on top of the same base within days or weeks, with low trial-and-error cost.

    But it also has a clear ceiling: post-training cannot conjure knowledge or capability that the base does not already contain. If a base’s underlying representations for a certain class of tasks are weak to begin with, no amount of alignment can do more than “reshape” it—it cannot “expand” it. So the premise of this route is that the base itself is already strong enough.

    2. Route Two: Tear it down and rebuild a large new base

    Sparse activation: MoE powering trillion-parameter models

    When a team decides to rebuild the base—especially at trillion-parameter scale—it almost inevitably moves toward a Mixture of Experts (MoE) architecture. The core idea of MoE is that the model has a very large total parameter count, but each forward pass activates only a subset of the expert networks.

    1. # MoE configuration illustration (fields are explanatory only, not a specific API)
    2. num_total_params: trillions # Total parameter count, determines knowledge capacity
    3. num_experts: N # Number of experts
    4. experts_per_token: top_k # Number of experts activated per token (k << N)
    5. capacity_factor: c # Load balancing and memory estimation

    This design of “large total parameters, small activated parameters” lets a model hold knowledge capacity close to that of a much larger dense model while keeping inference cost close to a small dense model. It is the mainstream engineering choice for large-scale bases today.

    Data engineering and training strategy

    Rebuilding a base is far more than “piling on parameters.” What really decides success or failure is:

    • Data quality and mixture: the domain distribution of pre-training corpora, deduplication quality, and the share of code and math data directly determine the model’s foundation for coding tasks.
    • Scaling-law guidance: under a given compute budget, sensibly balance “parameter count vs. training tokens” to avoid under-training or over-training.
    • Training stability: load balancing, expert routing, and communication overhead for trillion-scale MoE all require dedicated parallelism and optimization strategies, or you risk expert collapse or divergent training.

    Cost and payoff

    The cost of rebuilding a base is very high compute investment and a long iteration cycle. A single training run can consume enormous resources, and trial-and-error cost is far higher than with post-training. But the payoff is also direct: once the base’s knowledge capacity and underlying representations are substantially stronger, its capability ceiling is genuinely raised—not merely reshaped through alignment.

    3. Coding and agent capabilities: differentiated considerations for each route

    Coding and agent capabilities are the current focus of competition, and the two routes work differently on these two dimensions:

    • Coding capability depends both on the base’s “native understanding” of code syntax and heavily on verifiable-reward reinforcement during post-training. Rebuilding the base can raise the ceiling of underlying code representations; post-training can efficiently turn existing representations into “can write, can debug, can test” engineering ability.
    • Agent capabilities (multi-step planning, tool calling, long-horizon task execution) rely more on behavior alignment and trajectory training during post-training. The model needs to learn—in real or simulated interactive environments—to break down tasks, call tools, and correct itself based on feedback. This kind of capability is hard to acquire from pre-training alone.

    One key insight is that coding and agent capabilities are usually the joint product of “a strong base as the foundation plus post-training to shape it.” Relying on only one end is unlikely to reach the extreme.

    4. Engineering perspective: how to choose between the two routes

    For a concrete project, decide along these dimensions:

    Decision dimension Lean toward post-training extraction Lean toward rebuilding the base
    Compute budget Limited; needs fast iteration Ample; can accept long cycles
    Current base strength Already strong, with room to unlock Clearly insufficient, with capability gaps
    Timeline requirement Weekly delivery Monthly or longer
    Knowledge capacity needs Relatively focused tasks Broad coverage of knowledge and new capabilities

    A practical evaluation loop is: build a quantifiable evaluation set first, then decide whether to invest in rebuilding. The evaluation set should cover the target scenarios (e.g., code generation, unit-test pass rate, success rate of multi-step tool calls) and be benchmarked against public industry standards:

    1. # Route selection evaluation loop illustration
    2. def decide_route(base_model, budget, eval_suite):
    3. baseline = evaluate(base_model, eval_suite)
    4. if budget.can_afford_retraining and baseline.has_capability_gap():
    5. return "rebuild_base" # Sufficient compute and a capability gap exists
    6. else:
    7. return "post_train" # Base is strong enough; prioritize post-training extraction

    On the infrastructure side, whether it is distributed job orchestration for large-scale training, object storage and version management for training data, or containerized deployment and monitoring/alerting for inference services, teams can lean on the general capabilities of cloud platforms such as Baidu AI Cloud (e.g., container platforms, object storage, logging and monitoring services) to reduce the infrastructure burden and focus their energy on the model and the data themselves.

    5. Conclusion

    “Post-training extraction” and “retraining the base” are not mutually exclusive opposites, but two strategies under a trade-off among resources, time, and capability gaps:

    • When the base is strong enough and fast iteration is needed, post-training is the most cost-effective choice;
    • When the base has a fundamental capability gap and compute is ample, rebuilding a trillion-scale sparse base is what genuinely raises the real capability ceiling;
    • For complex capabilities like coding and agents, both a strong base and post-training to shape it are indispensable.

    For engineering teams, the most important thing is not to follow a “route myth,” but to first build a quantifiable evaluation system, drive decisions with data, and choose the path with the best return on investment under the given constraints.